AutoGPT - autonomous AI coming?

AutoGPT - autonomous AI coming?

"The most beautiful boy" in the AI ​​circle this week - AutoGPT

From the name, you can roughly see the main point: Auto , yes, yes, autonomous , autonomous artificial intelligence.

So what exactly is AutoGPT, and to what extent can it do it? Let's first watch a video together:

AutoGPT —— AgentGPT

The video content is provided by AgentGPT, and the effects of AgentGPT and AutoGPT are basically the same

In the video, I only set a goal of "I want wealth and freedom", AutoGPT!! #FA5151 independently formulated a complete plan and plan through "thinking"!! and personally "implemented" the plan.

The whole process is as follows:

1. Set a goal: I want financial freedom
2. AutoGPT disassembles the main task according to the set goals (the bottom layer is GPT)
   • Analyze the current financial situation and identify areas for optimization and improvement
   • Create a budget plan to increase savings and reduce unnecessary expenses
   • Explore investment opportunities and recommend solid solutions
3. Execute each main task separately and get the result
   • Executing the main task 1, as can be seen from the analysis results, is quite reasonable.
   • After executing Task 1, it added another subtask to itself based on the returned results: conducting market research to identify potential high-yield investment options. Equivalent to the returned result will be used as a new target
   • Execute the main task 2, and at the same time add additional subtasks according to the returned results of the task
   • …
4. When the main task is executed, additional subtasks will be executed.

The above process has been repeated, autonomously dismantling the target task, task acquisition plan (plan), and then dismantling independently, until the task is realized, and finally a very detailed plan for realizing wealth freedom will be obtained.

Wow, so scary. Isn't it a bit invincible? Wealth and freedom are so simple.

After watching the video, you must be wondering what AutoGPT is.

1. What is it?

AutoGPT can autonomously achieve any goal set by the user.

AutoGPT breaks down a goal and autonomously proposes an implementation plan.

AutoGPT is an open source project on Github, which has received a lot of attention, including Andrej Karpathy, the former AI director of Tesla and who just returned to OpenAI, also promoted it, and posted on twitter: AutoGPT will become the next frontier of prompt engineering .

The picture is taken from Andrej Karpathy's tweet

Some bigwigs even think that ChatGPT is outdated, and AutoGPT will be a new era.

Picture taken from Gregiseberg tweet

2. What can be done?

AutoGPT is taking AI circles by storm, and it's everywhere.

In addition to the goal plan seen at the beginning of the article, it also does a lot of things to improve productivity.

For example, to create a backend interface, the whole process includes creation, unit test, re-creation, re-test and then successful creation. The whole process is completely independent, and only one goal needs to be provided. **AutoGPT can not only complete the task, but also guarantee the quality** , let's say horror is not scary!

AutoGPT Generation API

The content of the video comes from tweets

Another example is that some big guys experimented only to let AutoGPT generate a website, and AutoGPT realized it in less than 3 minutes. The project is implemented using React and Tailwind CSS, the whole process is completely autonomous, only need to provide a target.

It seems that the programmer's dog head can't keep it.

AutoGPT generates web pages

The content of the video comes from tweets

3. What is the essence?

Here is a brief introduction to the principle of AutoGPT.

As Lior, a former AI researcher at the Mila Institute, said:
AutoGPT is equivalent to providing memory and subject for GPT-based models

Picture taken from Gregiseberg tweet

Specifically:

1. AutoGPT is based on GPT-4 and GPT-3.5, and the bottom layer is connected through API;
2. AutoGPT can iterate autonomously , and can obtain more accurate results by integrating prompt history;
3. AutoGPT has memory management , integrated database, so it can store history, implement context storage and make decision improvement based on it

So what is its essence? It's actually quite simple: LLM recursively calls

The picture is taken from Jay Hack tweet

In fact, this is not a patent of AutoGPT, a large number of tools are based on this idea, such as ViperGPT, SayCan, etc.

The picture is taken from Jay Hack tweet

The whole process is as follows:

1. Enter the target task
2. Send the target task to the execution client (Execution Agent), the bottom layer uses GPT-4, but this execution client is dedicated to understanding tasks, splitting tasks, and querying task results
3. Return the task list to the task creation client (Task Creation Agent). The bottom layer also uses GPT-4, and will create tasks independently based on the returned task results . This is where the autonomy is withdrawn .
4. Use the self-created task as the target task, and recursively call the implementation
5. All tasks will be prioritized based on GPT-4
6. The result returned by the task will be stored as a context for the next task creation

We simply understand it as a recursive call, @mathemagic1an understands it as a model stack , and the model is called downward, that is, the result is used as the input , and then called layer by layer .

4. How to try

1. Auto-GPT

• Project address: https://github.com/Torantulino/Auto-GPT
• The open source project can be directly cloned to deploy and run. The deployment steps are very detailed and simple in the README.md of the project.

2. AgentGPT

• Address: https://agentgpt.reworkd.ai/
• Experience address: https://github.com/reworkd/AgentGPT
• Open source project, can be deployed locally, basically the same effect as Auto-GPT

3. BabyAGI

• Project address: https://github.com/yoheinakajima/babyagi
• Open source project, can be deployed locally

5. Notes

1. Relatively high cost and difficult to control
   For a target, AutoGPT will recursively disassemble it and send it to GPT for a response. Moreover, the more general the target task is, the more subtasks will be required to solve this task, the greater the number of recursive layers will be, and the more tokens will be sent to GPT, and the corresponding cost will be more.

   Of course, you can try to limit the number of dialogue rounds to softly control it, but it is only a drop in the bucket, and the effect may not be satisfied after the limitation. After all, the realization principle of AutoGPT is to call recursively layer by layer to obtain the result.

2. The result is temporarily unable to interfere (training).
   The principle of AutoGPT is simple, but it is actually a recursive black box for us. There is currently no way to limit the call of each layer or customize the call.

6. Expectations

AutoGPT could be the next big thing in AI. There are more and more things that can be done and carried. I believe that there is still much room for imagination in GPT. We will continue to wait and see.

Sometimes I wonder what to do if AI replaces me, but if AI can really replace me, it's good to think about it!